This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2017.2762927, IEEE
Transactions on Power Delivery

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. , NO. , 2017 1

Assessing Gaussian Assumption of PMU

Measurement Error Using Field Data
Shaobu Wang, Member, IEEE, Junbo Zhao, Zhenyu Huang, Fellow, IEEE, Ruisheng Diao, Senior Member, IEEE

ƵƐϭ ƵƐϭ
Abstract—Gaussian PMU measurement error has been as-
sumed for many power system applications, such as state estima-
tion, oscillatory modes monitoring, voltage stability analysis, to
cite a few. This paper proposes a simple yet effective approach ƌĞĂŬĞƌϭ
to assess this assumption by using the stability property of a
probability distribution and the concept of redundant measure-
ƌĞĂŬĞƌϮ
ment. Extensive results using field PMU data from WECC system
reveal that the Gaussian assumption is questionable.
Index Terms—Non-Gaussian distribution, PMU measurement, ƌĞĂŬĞƌϯ
stable distribution, power system operation, state estimation.
ƵƐϮ ƵƐϮ
I. I NTRODUCTION
ƌĞĂŬĞƌĂŶĚĂŚĂůĨ ŽƵďůĞͲďƵƐͲĚŽƵďůĞͲďƌĞĂŬĞƌ

W ITH the wide-area deployment of phasor measurement

units (PMUs), many power system online monitoring
and control applications become possible, including dynamic Fig. 1.
ĐŽŶĨŝŐƵƌĂƚŝŽŶ ĐŽŶĨŝŐƵƌĂƚŝŽŶ

Different Configurations of Circuit Breakers.

state estimation, oscillatory modes monitoring, voltage sta- To analyze error statistics, the first step is to extract them
bility analysis, to name a few [1]–[3]. However, Gaussian from measurements. However, as practical power systems are
measurement error is usually assumed when conducting those typically non-stationary, it is very challenging for traditional
applications. It should be noted that if Gaussian assumption is approaches, such as finite impulse response (FIR) digital
violated, the obtained results could be misleading and harmful low pass filters, median filter [4]–[6]. An alternative way
for power system security. to extract error from field PMU measurements is through a
In this letter, we propose a simple yet effective approach to high precision calibrator, where same inputs are fed into the
assess this assumption by using the field PMU measurements calibrator and the PMU devices under test, then the differences
from the Western Electricity Coordinating Council (WECC) of their outputs are taken as the measurement error. Although
system. We utilize the concept of redundant measurement to this is the best way to obtain almost true error statistics, it
construct an error vector for Gaussian assumption assessment. is not economically feasible as the high precision calibrator
The latter is checked by the stability property of a probability is very expensive. In this paper, we propose a redundant
distribution and the Shapiro-Wilks test. Extensive results show measurement-based approach to check the statistical property
that PMU measurement error potentially follows thick-tailed of the measurement error.
non-Gaussian probability distributions.
II. P ROBLEM F ORMULATION III. P ROPOSED A PPROACH
Definition 1. PMU measurement error includes the absolute A. Concept of Redundant Measurements
error (or bias), the relative error and noise that is induced by Redundant measurements are often set up to guarantee
instruments, communication channels, etc. Bias is a constant observability for different network topologies or operating
instead of a random variable, and therefore it does not modes. For example, Fig. 1 shows the breaker-and-half
affect the distribution of the measurement error. It should be bus/switching configuration and double-bus-double-breaker
noted that measurement noise sometimes has been used with configuration cases in the WECC system. In order to ensure
measurement error interchangeably in the literature. However, the observability of buses 1 and 2 for the breaker-and-half
they are in fact different and measurement error is more bus/switching configuration case when any one of the three
general. In addition, it is the measurement error that we care breakers opens, two PMUs are installed, one at each bus.
most when implementing measurement-based applications. PMUs measure the bus voltage phasor as well as the current
phasors of the lines that are adjacent to that bus. Note that
This work is partially sponsored by U.S. Department of Energy through we get the information of the breaker status from SCADA
its Advanced Grid Modeling program and its ASCR Applied Mathmematics data. When all the breakers get closed, the metered voltage
program, and U.S. National Science Foundation Award ECCS-1711191.
Shaobu Wang, Zhenyu Huang and Ruisheng Diao are with Pa- magnitudes and angles or line current phasors obtained by
cific Northwest National Laboratory, Richland, WA 99352, USA (email: these two PMU devices should be identical in absence of
shaob.wang@pnnl.gov, zhenyu.huang@pnnl.gov, ruisheng.diao@pnnl.gov). errors. Then, we say that PMUs installed at buses 1 and 2
Junbo Zhao is with the Bradley Department of Electrical and Computer
Engineering, Virginia Polytechnic Institute and State University, Falls Church, form a redundant case. Furthermore, if these two PMU devices
VA 22043, USA (e-mail: zjunbo@vt.edu). are from the same vendor and operate with almost the same

0885-8977 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2017.2762927, IEEE
Transactions on Power Delivery

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. , NO. , 2017 2

environmental factors, we would expect that the error statistics

of them should be approximately the same.
B. Assessment of Gaussian Assumption
Let S be the set of redundant measurements, for i, j ∈ S,
define xi = x + ei and xj = x + ej as the measured
values provided by PMUs, where x is the true value; e i
and ej are measurement errors associated with ith and jth
PMUs. Note that xi and xj come from different PMU devices,
they are considered as independent. Furthermore, to extract
measurement error from the redundant set defined in Section
(a)
III-A, we define Δx ij = xi − xj = ei − ej . Subsequently,
we have the following theorem:
Theorem 1. A necessary condition that both e i and ej follow
a Gaussian distribution is that Δx ij follows a Gaussian
distribution.
Proof: We will prove this theorem by contradiction.
Case 1: assume that both e i and ej follow a Gaussian dis-
tribution, then by the stability property of the normal/Gaussian
distribution, it is straightforward to verify that Δx ij follows
a Gaussian distribution.
Case 2: assume ei (or ej ) follows a Gaussian distribution
(b)
while ej (or ei ) has a non-Gaussian distribution, Δx ij must
follow a non-Gaussian distribution. This is because if Δx ij
is Gaussian, by subtracting e i (or ej ) from Δxij and using
the stability property of the Gaussian distribution, e j (or
ei ) must follow a Gaussian distribution. This contradicts the
assumption.
Case 3: assume both ei and ej follow a non-Gaussian distri-
bution, Δxij follows a Gaussian or non-Gaussian distribution.
From cases 1-3, the Theorem follows.
Remark 1: according to the central limit theorem, it is
true that the difference between two non-Gaussian random
(c)
variables can yield a Gaussian distribution as the number of
measurements tends to infinity. However, this can hardly hold
true for practical power systems. This is because the measure-
ment error of PMU depends on the operating conditions of the
system and many environmental factors, yielding time varying
error statistics. As a result, the independent and identically
distributed random variable assumption of the central limit
theorem does not hold true.
Remark 2: in our studies, two similar PMUs (from the same
vendor) measure the same quantities and are subject to same
impacts of environmental factors. We can anticipate that the (d)
error statistics of these two PMU devices are quite similar.
Thus, according to the stability property of a probability
distribution and Theorem 1, if we find that Δx ij follows a
Gaussian distribution, it is likely that both e i (or ej ) follows
a Gaussian distribution; by contrast, if Δx ij follows a non-
Gaussian distribution, it is of very high probability that both
ei (or ej ) follows a non-Gaussian distribution.
Following Theorem 1, validation of the normality of Δx ij
becomes the key step. To do that, several widely used sta-
tistical tests can be applied, including W/S test, Jarque-Bera (e)
test, Shapiro-Wilks test, Kolmogorov-Smirnov test [7]. Among Fig. 2. Results of the Shapiro-Wilks test for bus set 1 with varying
them, Shapiro-Wilks test [8] is chosen in this paper because number of measurement samples; 500, 1000, 1500, 3000 and 6000 for (a)-
(e), respectively; the frequency for y-axis represents the number of times the
of its well demonstrated performance in many applications. It measurement error fall into each bin.
uses the null hypothesis principle to check whether a sample

0885-8977 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRD.2017.2762927, IEEE
Transactions on Power Delivery
IEEE TRANSACTIONS ON POWER DELIVERY, VOL. , NO. , 2017
Fig. 3. Results of the Shapiro-Wilks test for bus set 2.
Fig. 4. Estimated probability distributions of the error terms for bus set 2.
series is from a normally distributed population. Note that the
null hypothesis corresponds to the Gaussian assumption. In
the test, two parameters, namely p-value and α, need to be
studied. Parameter α represents the significance level of the
statistical test and is usually set according to the confidence
level of significance. Note that in many existing statistical
test approaches for validating normality/Gaussian assumption,
anomaly or outlier detection, 95% is the value widely adopted
for confidence level of significance. As a result, α is set to be
1 − 0.95 = 0.05. In this case, if the Shapiro-Wilks test returns
a p value and p ≥ α, we conclude that the random variable
follows the Gaussian distribution. The Shapiro-Wilks test can
be implemented in Matlab by calling the function SWTEST.
IV. T EST R ESULTS WITH F IELD DATA
To test the effectiveness of the proposed approach, field
PMU measurements from the WECC system are used. Note
that there are many redundant measurement cases in this
system, from which nine sets of PMU redundant measurements
of 18 buses are investigated. The scan rate of the PMUs is 30
samples/s.
A. Choice of PMU Data for Testing
For all hypothesis tests, two important factors need to be
investigated: sample size, a number that is directly related
to the confidence of the test and the expected false positive
rate, and the test power, which represents the chance of false
negatives. When the sample size is small and you insist on it
with high confidence, the test power gets worse. By contrast,
with larger sample size, the test power gets better. Note that
many statistical tests will become inaccurate if the sample
size is very large (larger than 6000 for Shapiro-Wilks test
for example). This is because according to the central limit
theorem, the sampling distribution tends to be normal if the
number of samples is very large, regardless of the shape of
the data. In this paper, we have tested the cases with varying
number of measurement samples for bus set 1 and the results
are shown in Fig. 2, where p-value is shown at the top of each
0885-8977 (c) 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
3
figure. It is clear that the measurement errors of both voltage
magnitudes and angles do not follow a Gaussian distribution
as their p values are close to 0. In addition, it is found that the
shapes of the histograms remain unchanged for a large number
of samples (see the cases with sample sizes 3000 and 6000
for example). Thus, the number of samples between 3000 and
6000 is a good choice.
B. Results for Multiple Bus Sets and Recommendations
Tests have been carried out on other 8 bus sets that form
measurement redundant cases as well. It is found that none
of the sets has a normally distributed PMU measurement
error. Instead, they follow a non-Gaussian distribution with
long tails, such as student-t, logistic, Laplace distributions.
It should be emphasized that due to different system oper-
ating conditions, aging process of the PTs and CTs, varying
communication channel noises, PMU measurement error can
change from time to time. It is thus difficult to suggest
a specific probability distribution of the measurement error.
On the other hand, it is worth pointing out that in many
signal processing problems, student-t and Laplace distributions
are the two most widely used ones to model thick-tailed
distributions, which can be used to simulate realistic PMU
measurement error. For example, the Shapiro-Wilks test results
of bus set 2 are displayed in Fig. 3. In the meantime, the
estimated probability distributions of the error terms shown
in Fig. 3 are displayed in Fig. 4. From these two figures, we
observe that the measurement error follows a non-Gaussian
distribution with long tails, such as student-t and logistic
distributions. Note that these two most probable distributions
are selected according to the significance confidence values of
the probability distribution fitting test.
V. C ONCLUSION
This letter proposes a simple yet effective approach to
investigate the measurement error of the field PMU data. It is
found that the realistic PMU measurement errors are unlikely
to follow a Gaussian distribution. In future work, we will
investigate the impacts of non-Gaussian measurement error on
power dynamic state estimation, oscillatory modes monitoring,
etc.
R EFERENCES
[1] J. B. Zhao, M. Netto, L. Mili, “A robust iterated extended Kalman filter
for power system dynamic state estimation”, IEEE Trans. Power Syst.,
Vol. 32, no. 4, pp. 3205-3216, 2017.
[2] B. Milosevic, M. Begovic, “Voltage stability protection and control using
wide-area network of phasor measurements,” IEEE Trans. Power Syst.,
vol. 18, no. 1, pp.121–127, 2003.
[3] Z. Huang, P. Du, et. al, “Generator dynamic model validation and param-
eter calibration using phasor measurements at the point of connection,”
IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1939–1949, 2013.
[4] M. Brown, M. Biswal, et. al, “Characterizing and quantifying noise in
PMU data,” In Proceedings IEEE PES General Meeting, pp. 1-5, 2016.
[5] J. Tate, T. Overbye, “Extracting steady state values from phasor mea-
surement unit data using FIR and median filters,” in IEEE PES Power
Systems Conference and Exposition, pp.1-8, 2009.
[6] D. L. Donoho, “De-noising by soft-thresholding,” IEEE Trans. on Inf.
Theory, Vol. 42, no. 3, pp. 613-627, May, 1995.
[7] H. J. Thode. Testing for normality. New York: Marcel Dekker, 2002.
[8] S. Shapiro, M. Wilk, “An analysis of variance test for normality
(complete samples),” Biometrika, Vol. 52, no. 3, pp. 591-611, 1965.